Newsgroups: comp.parallel.mpi
From: gustav@arp.anu.edu.au (Zdzislaw Meglicki)
Subject: Re: Heterogenous MPI
Organization: Centre for Information Science Research, The Australian National University
Date: 7 Mar 1995 00:19:14 GMT
Message-ID: <3jg8q2$ffk@manuel.anu.edu.au>

Although MPI provides adequate hook-ups to run on a heterogeneous
farm, I vividly remember a comment by Prof. Tanenbaum during HPDC Conference
in Spokane, 1993, who said that "heterogeneous distributed computing is
bad research". What he meant was not that you shouldn't use a heterogeneous
farm if you have one, but that given a choice between a homogeneous system
and a heterogeneous one, you'll be always better off using the former.
The problems posed by converting data formats and varied CPU performance
of such systems effectively degrades performance of your distributed
applications. Naturally this comment applies to scientific and engineering
computations where what you want first and foremost is speed. 

In business computing or factory floor wiring problems are different.
There you may want to program a system of functionally different robots
utilising different CPUs, yet they should all cooperate on a common task, 
for example assembling a car. Speed is a secondary issue in this context. 
Even a fastest mechanical system is orders of magnitude slower than 
electronics. You can just as well do it all in Lisp and save yourself 
the indignity of programming in C!

-- 
 Zdzislaw Meglicki, Zdzislaw.Meglicki@cisr.anu.edu.au,
 Parallel Computing Research Facility, CISR,
 The Australian National University, Canberra, A.C.T., 0200,
 Australia, fax: +61-6-249-0747, tel: +61-6-249-0158

